Impacts of Frequent Itemset Hiding Algorithms on Privacy Preserving Data Mining

Abstract

The invincible growing of computer capabilities and collection of large amounts of data in recent years, make data mining a popular analysis tool. Association rules (frequent itemsets), classification and clustering are main methods used in data mining research. The first part of this thesis is implementation and comparison of two frequent itemset mining algorithms that work without candidate itemset generation: Matrix Apriori and FP-Growth. Comparison of these algorithms revealed that Matrix Apriori has higher performance with its faster data structure. One of the great challenges of data mining is finding hidden patterns without violating data owners. privacy. Privacy preserving data mining came into prominence as a solution. In the second study of the thesis, Matrix Apriori algorithm is modified and a frequent itemset hiding framework is developed. Four frequent itemset hiding algorithms are proposed such that: i) all versions work without pre-mining so privacy breech caused by the knowledge obtained by finding frequent itemsets is prevented in advance, ii) efficiency is increased since no pre-mining is required, iii) supports are found during hiding process and at the end sanitized dataset and frequent itemsets of this dataset are given as outputs so no post-mining is required, iv) the heuristics use pattern lengths rather than transaction lengths eliminating the possibility of distorting more valuable data.